类型错误：数组必须具有一致类型的分配

Question

从here的后续行动，我有类似下面的代码：

@jit(float_[:,:,:](float_[:,:], int_[:], int_)) 
def train_function(X, y, H): 
    # do lots of stuff, including setting the arrays g and g_per_round like this: 
    g = np.zeros((no_features, no_classes)) 
    g_per_round = np.zeros((H, no_features, no_classes)) 

    # do more stuff, then:  
     g_h = None 
     j = 0 
     print "Calculating regression coefficients per class. .." 
     # building the parameters per j class 
     for y1_w in zip(z.T, weights.T): 
      y1, w = y1_w 
      temp_g = sm.WLS(y1, X, w).fit() # Step 2(a)(ii) 
      if g_h is None: # sometimes g *is* None, and that's fine 
        g_h = temp_g.params # this is an array of floats 
      else: 
        g_h = np.c_[g_h, temp_g.params] 
      j = j + 1 

     if np.allclose(g,0) or g is None: 
      g = g_h 
     else:    
      g = g + g_h 

    # do lots more stuff, then finally: 
    return g_per_round 

class GentleBoostC(object): 
    # init functions and stuff 
    def train(self, X, y, H): 
     self.g_per_round = train_function(X, y, H)  

现在我发现了以下错误：

@jit(float_[:,:,:](float_[:,:], int_[:], int_)) 
more lines, etc etc etc, last few lines: 
    unresolved_types, var_name) 
    File "C:\Users\app\Anaconda\lib\site-packages\numba\typesystem\ssatypes.py", line 767, in promote_arrays 
    assert_equal(non_array_types[0]) 
    File "C:\Users\app\Anaconda\lib\site-packages\numba\typesystem\ssatypes.py", line 764, in assert_equal 
    var_name, result_type, other_type)) 
TypeError: Arrays must have consistent types in assignment for variable 'g': 'float64[:, :]' and 'none' 

我在尝试添加@jit以加速我的代码之前，实际上没有任何问题。

Answer 1

问题是Numba推断g_h为NoneType;它初始化为载体，将正确编译：

g_h = np.zeroes((H, no_features, no_classes)) 

Answer 2

的问题是，numba无法知道g_h不会None当它最终分配到g因为g_h类型取决于运行时的流量控制。换句话说，如果g_h可能永远不是一个float64，那么它必须假定有时不是。

这是一个documented limitation of numba和类型推断系统的一般的限制：

However, there are some restrictions, namely that variables must have a unifyable type at control flow merge points. For example, the following code will not compile:

@jit def incompatible_types(arg): 
    if arg > 10: 
     x = "hello" 
    else: 
     x = 1 

    return x  # ERROR! Inconsistent type for x! 

的解决方案是初始化g_h到兼容类型，而不是= None。

Numba的类型推断其实是相当聪明，所以你可以不用在很多情况下问题，只要类型可以返回之前被统一混合在一个特定的局部变量的类型。阅读Numba documentation on types了解更多信息。